Last Updated on July 7, 2026 by Taya Ziv
For two years, DeepSeek was the startup that made Silicon Valley uncomfortable for all the right reasons. A small team in Hangzhou, funded entirely by a hedge fund’s profits, building models that matched or beat the ones costing tens of billions to develop. No venture capital. No board seats traded for capital. No investor decks. Just a quant trader named Liang Wenfeng who decided that artificial intelligence was a more interesting problem than beating the stock market, and proved it by spending a fraction of what everyone else was spending.
That story just ended.
According to reports from Bloomberg and the Financial Times, China’s Integrated Circuit Industry Investment Fund, known as the Big Fund, is in talks to lead DeepSeek’s first outside funding round at a valuation north of $45 billion. The round is expected to bring in $3 to $4 billion. For a company that turned down every VC firm, every tech giant, and every state fund that came knocking for two straight years, this isn’t just a change of strategy. It’s a signal that something fundamental shifted.
And what shifted matters a lot more than the number on the term sheet.
The Self-Funded Miracle That Couldn’t Last
Let me rewind for a second, because the DeepSeek origin story is genuinely one of the weirdest in tech.
Liang Wenfeng co-founded High-Flyer, a quantitative hedge fund, in 2016. The firm used AI for algorithmic trading and got very good at it. By 2021, High-Flyer was running entirely on AI-driven strategies. In 2023, Liang spun off the AI research lab into a separate company: DeepSeek. The funding source? High-Flyer’s trading profits. The hedge fund reportedly posted a 56.6% return in 2025 alone.
So DeepSeek had something almost no AI startup has: a parent company printing money, no outside investors demanding growth metrics, and a founder who genuinely didn’t care about fundraising optics. The result was a company that could take big technical bets without anyone asking about the unit economics on slide 14.
And those bets paid off spectacularly. DeepSeek’s models, especially V3 and the recently released V4, dropped the cost of frontier AI by roughly 90% compared to what OpenAI and Anthropic charge. V4 Pro matches GPT-5.5 and Claude Opus 4.7 on most benchmarks while costing about $3.48 per million output tokens. OpenAI charges $30 for the same work. The models are open-weight, self-hostable, and free to modify.
This was the narrative that terrified the American AI industry. A company spending less, shipping faster, and giving it away for free. The ultimate price disruptor in a market where everyone else was raising billions.
So why, after two years of saying no, did they say yes?
Three Theories About What Changed
Theory one: the compute wall hit. DeepSeek built V3 on what’s estimated to be around 10,000 Nvidia H800 GPUs, hardware it acquired before the most restrictive US export controls kicked in. V4 is a 1.6-trillion-parameter model. The next generation will be bigger. At some point, even a hedge fund printing 56% annual returns can’t keep pace with the compute demands of frontier model development. The Big Fund’s $3-4 billion isn’t just money. It’s access to a government pipeline for chips.
Theory two: the sanctions made state connections essential. The US has been tightening chip export controls to China for three years. DeepSeek’s entire competitive advantage depends on getting enough compute to train models. China has been building its own chip ecosystem, and DeepSeek’s V4 is already being optimized for Huawei’s Ascend chips. But getting priority access to domestic chips means being on the right side of the government. The Big Fund is literally the vehicle Beijing uses to control who gets semiconductors and who doesn’t.
Theory three: Beijing made an offer that was also a strategic claim. The Big Fund doesn’t invest in software companies. It invests in semiconductor infrastructure. Its portfolio is chip fabs, memory manufacturers, and packaging companies. When it leads a round in an AI company, it’s not making a financial bet. It’s making a geopolitical statement. DeepSeek is being elevated from “interesting Chinese AI startup” to “strategic national asset.” And that status comes with expectations.
My guess? All three are true, and they feed each other. The compute needs created the opening. The sanctions made state backing valuable. And Beijing saw a chance to claim the company that embarrassed Silicon Valley as its own.
What “Open Source” Means When Your Investor Is a Government
Here’s the question nobody is asking yet, and founders everywhere should be thinking about.
DeepSeek has been the poster child for open-source AI. Its models are available for anyone to download, modify, and deploy. That openness was a competitive weapon. It forced OpenAI and Anthropic to defend their pricing. It gave thousands of startups around the world access to near-frontier AI capabilities for free. It accelerated the commoditization of the model layer in ways that benefit everyone building applications on top of it.
But open source backed by a hedge fund is different from open source backed by the Chinese government’s semiconductor sovereignty fund.
I’m not saying DeepSeek will suddenly close its models. That would be strategically stupid and they know it. The open-weight release is what makes DeepSeek important. Without it, they’re just another Chinese AI lab.
But the incentive structure just changed. The Big Fund’s portfolio companies serve Beijing’s industrial policy goals. The fund was literally created to reduce China’s dependence on foreign semiconductors. When that fund puts $3-4 billion into an AI company, the company becomes part of a national strategy, whether the founders intended that or not.
For every startup currently building on top of DeepSeek models, this is worth watching closely. Your foundation layer just became part of a geopolitical chess game. It might not change anything tomorrow. But the risk profile is different now.
What This Tells Founders About the AI Market
Forget China for a minute. The deeper signal here is about the economics of building frontier AI.
DeepSeek proved that you can build world-class models without spending like OpenAI. That was real. The efficiency gains were documented and reproducible. V4’s mixture-of-experts architecture activates only 49 billion of its 1.6 trillion parameters per query, which is why the costs are so low.
But even DeepSeek, the most capital-efficient AI company on earth, eventually needed billions. If the leanest operation in frontier AI can’t sustain itself on internal funding, what does that say about every other AI company that’s been telling investors they can reach profitability?
The math is simple and a little terrifying. SoftBank just borrowed $40 billion for AI infrastructure. OpenAI is approaching $25 billion in annualized revenue and still burning cash. Anthropic is at $19 billion and raising more. The foundation model layer is a money furnace. Even the company that proved it could run on less fuel eventually needed more fuel.
If you’re a startup building in the application layer, this is actually good news. The model layer is commoditizing, the price war is real, and the companies fighting it need so much capital that they’ll never have the focus or the margins to compete with you on vertical applications. Build on top of the commodity. Let the model makers bleed.
But if you’re a startup building at the model layer and you think you can out-efficiency DeepSeek? The company that pioneered that playbook just admitted it wasn’t enough. That should recalibrate your expectations.
The $45 Billion Question
DeepSeek’s valuation jumped from $10 billion to $45 billion in about three weeks. That’s a 4.5x increase driven not by revenue metrics, but by which government decided to invest. No ARR was disclosed. No customer count. No margin data. Just a state fund writing a check and a valuation multiplying overnight.
That’s not how startup valuations are supposed to work. But in AI, in 2026, this is how they do work. The geopolitical value of frontier AI research now exceeds the commercial value in certain contexts. DeepSeek isn’t being priced on what it earns. It’s being priced on what it represents.
For two years, DeepSeek represented the idea that brains beat budgets. That small teams with clever architectures could outrun the giants. And they did.
Now DeepSeek represents something else: the moment when even the disruptor needed the machine. The lean rebel became a state champion. The open-source idealist became a geopolitical asset.
That doesn’t make their models less impressive. V4 is still remarkable technology. The efficiency innovations are still real. The cost disruption is still happening. But the company behind them is no longer the scrappy outsider. It’s sitting at the table with the biggest fund in Chinese industrial policy, and the conversation isn’t about benchmarks anymore.
The AI race isn’t just a technology race. It’s a sovereignty race, the newest front in the AI infrastructure wars. And today, Beijing made its loudest move yet.
